VIENNA INSTITUTE OF DEMOGRAPHY WORKING PAPERS - Oeaw.ac.at

The reconstruction follows the principle of back projections. Like in a forwardprojection, back projections require the availability of (ad A) a base-year dataset, in thiscase the most recent valid data on country-specific educational attainment (see Section 3.1).We assembled data from multiple sources and years. The most accurate dataset was chosenbased on several criteria detailed in Section 3.1.1. The dataset was further adjusted andharmonised to fit standard education categories (see Section 3.1.2). A feasible base-yeardataset as primary input data determines to an important extent the quality of thereconstruction exercise.The base year is used (ad C) in the reconstruction model to go back in time using thegeneral principles that (1) education is predominantly acquired at young ages, and (2) thateducation is acquired in a unidirectional mode. Individuals can only add skills andeducational levels until reaching their personal highest educational level, which becomes afixed attribute for the remaining life. Already achieved educational levels cannot bereversed. This allows to follow the educational progress of an individual back in time alongcohort lines (Goujon et al. 2016; Speringer, Goujon and Jurasszovich 2018).Therefore, a person i should have the same education at time t and at time t-5 in theperiod after leaving school/university until death. This is also valid at the aggregate level:the share of the population by level of educational attainment can be back-projected in timealong cohort lines. In country j, the share s of population with education k should be similarat time t and at time t-5. However, two factors could upset this equivalence and affect theeducational distribution:› Mortality: if the probability of surviving between time t-5 and time t differs by levelof education k;› Migration: if the probability of (in- or out-)migration between time t-5 and time tdiffers by level of education k;To account for education-specific differentials in mortality, we used information onmortality rates from the United Nations Population Division (2017) and applied standardmortality differentials by levels of education as developed by Lutz, Butz and KC (2014). Thedifferentials vary between genders and across the reconstruction period (see Sections 3.2.2.1and 3.3.1). It is not possible to account for education-specific differentials in migration sincethese cannot be standardised in the same way as for mortality. However, by using historicaldata points as mentioned below, we expect that the effect of migration on the educationstructure will be taken care of.While the above reconstruction procedure is valid for the out-of-school/universitypopulation, this is not necessarily the case for the schooling/studying age groups of 15 to 34years. 2 Therefore we developed a procedure to calculate over the reconstruction periodDisaggregation by education starts at age 15. Based on evidence, we consider that most educationtransitions happen by the age of 35.26

country-specific education transition rates that are applied to the population below the ageof 35 years (see Section 3.3.2).It is worth remembering that the reconstruction is done for the education distribution(in %) by age and sex, from 1950 to 2015. These shares are applied to the population by ageand sex for the same period as estimated for each country of the world by the UnitedNations Population Division (2017). The reconstruction principles are to a large extent thesame as those used in the WIC 2014 dataset, though the periodical adaptation of educationspecific mortality rates is a novelty in this version, and the estimation of educationtransitions is carried out in a different way in WIC 2018. Another innovation in WIC 2018is the use of information contained (ad B) in historical datasets on education (and literacy)when available for the period 1950 to 2015 (see Section 3.2). The collected historical datasetswere checked for accuracy and usability (see Section 3.2.1). They are used for two mainpurposes. First, they replaced missing education data when age groups are depletedthrough the reconstruction. For instance, the share of the population by levels of educationfor the age group 100 in 2010 will be used to calculate the share of the population in agegroup 95–99 in 2005. However, the share of the 100 population in 2005 will most likely notbe available and has to be estimated. In this way we incorporate historical cohortinformation on the educational composition in the reconstruction process (see Section3.2.2.3). The second purpose of historical data is to check the accuracy of the reconstructionoutput at each step. 3 In a final step, (ad D) the quality of the reconstructed WIC 2018 datasetis assessed and compared with alternative reconstruction exercises.3Data and Methodology3.13.1.1Assembling the Base-year DataBase-year Data Collection: Coverage & Data SourcesThe base-year data on population by age, sex and educational attainment described belowserves as a starting point for both the multistate population projections (Lutz et al. 2018)and the update of the reconstruction that is documented in this report. The previousreconstruction exercise, WIC 2014, was based on the collected and harmonised base-yeardata for 171 countries (Bauer et al. 2012). However, this dataset relied mostly on theinformation from the 2000 census round, as the more recent data from the 2010 censusround were not yet accessible at the time of the data collection. We also took the opportunityto fill data gaps and improve the quality of the dataset. 4In the WIC 2014 we were also using the historical data, but only to assess reconstruction results andnot as marker points in the reconstruction model itself.4 Furthermore, we have implemented the new ISCED 2011 classification (UNESCO 2012) todisaggregate the post-secondary education category into three education categories of highereducation for countries with available data (mostly OECD countries). We document this work in37

The starting point for the update of the educational attainment by age and sex was theprevious dataset that included information for 171 countries (Bauer et al. 2012). Aspreviously, the aim was to collect as recent as possible information on population by age,sex and educational attainment for 201 countries listed in the 2017 Revision of the UNWorld Population Prospects. In terms of geographical coverage, it was possible to collectand harmonise data for 185 countries (92% of all countries), covering 99% of the world’spopulation. This makes this dataset the most comprehensive internationally. 5In comparison to the WIC 2014 dataset, the country coverage has improved from 171countries covering 88% of all countries and 97.4% of the world’s population (Goujon et al.2016). Table 1 summarises the data availability and lists newly added countries and thosewith missing education data. The countries with missing education data are notreconstructed as they are lacking the base-year information.6Some of the 16 new countries that were added to the dataset came to existence due topolitical changes—e.g. Sudan split into Sudan and South Sudan, while several others wereadded to the data collection through population growth as they have exceeded or cameclose to the 100,000 population size threshold, e.g. Kiribati. To highlight some modificationsin comparison to the WIC 2014 dataset: in the WIC 2018 dataset the Netherlands Antilleswere replaced by Curaçao, the only one of the now independent entities with a populationexceeding the 100,000 threshold. In addition, Taiwan, which had been included underChina in WIC 2014, was added as a separate entity.Looking at the continents, Table 1 shows that the data coverage increased most forOceania (up from covering 76% to 80% of the region’s population), followed by Africa,(from 96% to 99% coverage) and Asia (97% to 99%) and with small improvements for LatinAmerica. The coverage did not change for the remaining regions. In Africa, some sizeablecountries such as Angola and Botswana were added. New datasets also became availablefor Afghanistan, Oman, North Korea, Sri Lanka and Yemen.Annex B as it is mostly relevant for the projections (to 2100—see Lutz et al. 2018) and not for theback-projections.5 For comparison, the Barro and Lee dataset covers in its latest version 146 counties (Barro and Lee2015).6 However, they are used in the projections, and their educational compositions is then imputedusing proxy countries from the region (see Annex A).8

Table 1. Country coverage of the updated WIC 2018 dataset grouped by UN regionsUN regionEuropeAsiaCountriesCountries Population Countries with New countriesAllwithmissingcovered incovered coveredcountries educationeducation data WIC 2018*(in %)(in stanAfghanistan,North Korea,Oman,SriLanka, Taiwan,YemenAngola,Botswana,South Sudan,Sudan, Togo-99.996.1Africa575087.798.6Djibouti,Eritrea, Libya,Mauritania,Mayotte,SeychellesWestern SaharaNorthernAmerica22100.0100.0-Latin America da, U.S.Virgin IslandsFiji,Kiribati,PapuaNew Micronesia,Guinea, Guam SolomonIslands--* compared to WIC 2014Due to lack of data, insufficient level of detail of the published educational data or dueto other data quality issues, information for 16 countries out of 201 countries is missing(down from 24 in WIC 2014). Among the populous countries with missing education dataare Uzbekistan (the last full census was implemented in 1989 and the only data availablecomes from a Demographic and Health Survey (DHS) in 1996) and Papua New Guineawhere our request for data was not answered.Besides data for 16 new countries (blue in Figure 1), information for 101 countries wasupdated (orange in Figure 1), either in terms of more recent census or survey, or in termsof data quality. For 68 countries, the same data source as in the previous baseline (WIC2014) is used (grey in Figure 1). The WIC 2018 dataset has information for 112 countrieswith data pertaining to 2010–2015 (up from 12 in WIC 2014 dataset). Good quality data arestill hard to get or for many countries in the Middle East, Africa and Central Asia, in many9

cases due to political unrest and violent conflicts. For example, it is still necessary to rely ondata sources from before the year 2000 for the Central African Republic (DHS 1995),Pakistan (1998 census as the 2017 census in not yet available) and Turkmenistan (last censusin 1995).Figure 1. Update of the WIC 2018 educational data (Source: authors’ own calculation)An increase in data quality was achieved by avoiding pre-compiled data sourcesprovided by UNESCO or other agencies as these data compilations often suffer from biases.Thus, for instance, census data were mostly collected from National Statistical Institutes(NSO) and in the greatest possible level of detail when it comes to age and educationcategories, avoiding pre-defined aggregate education categories that often do notcorrespond to the ISCED 2011 definitions, and are thus of limited use or require additionaladjustments.In general, register and census data tend to provide more complete (for example captureall age groups not only persons in reproductive or economically active ages) and reliableinformation than most surveys. 7 For countries without census data, we used the new wavesof surveys such as DHS or Multiple Indicator Cluster Survey (MICS), when available.Figure 2 and Figure 3 show the change in the type of primary data source for thecountry-specific base-year data between the WIC 2018 and the WIC 2014 datasets. The latteris composed of more register or census data for more countries than the former. The WICThis may not be the case in all countries as it has been documented that some census results wereflawed, e.g. Nigeria 2006 census or Chile 2010 census. In such cases, we have either used an oldercensus or alternative data, such as DHS and national households surveys. The selection of the finaldata source is subject to careful time-series analysis and cohort checks of the educational attainmentdata.710

2018 dataset includes census data for 126 countries from the 2010 census round, and for 20countries from the 2000 census round (Figure 3). For five countries (Austria, Switzerland,Estonia, Norway and Sweden), the most up-to-date register data at the time of datacollection was used. We replaced the previous WIC 2014 survey data with newerinformation (mostly from DHS) for 25 countries. If register or census data were notavailable, reliable or of sufficient quality, representative sample surveys (6 countries), DHS(19 countries) or other household surveys (9 countries) have been used.Figure 2. Data types used in WIC 2018 dataset for 185 countries (Source: authors’ owncalculation)11

Figure 3. Data types used in WIC 2014 dataset for 171 countries (Source: author’s owncalculation)Countries collect and publish education data according to their definitions and needs.Low and Middle-income Countries (LMICs) that strive mostly to increase enrolment inprimary and secondary, along the recommendations of international goals (such as theMillennium Development Goals or the Sustainable Development Goals) often publishdetailed statistics by level and grade completed. It was possible to collect all main sixeducational categories for 150 countries (Case 1 countries in Table 2).In contrast, High-income Countries (HICs) often focus the data collection on the higherend of the educational spectrum as virtually the entire working-age population hasachieved lower secondary education. 8 However, this hides the progress in educationalattainment that many societies have made during the past century and that should still bevisible among older age groups. Such information is important when assessing theeducational compositions of whole populations, including older age groups who havestudied under very different educational systems and in times when compulsory educationwas set at lower levels 9 and not strictly implemented. To fill those data gaps, alternative orAs foreign-born populations increase in many economically advanced countries there isnonetheless a need for educational data among the immigrants who may have lower educationalattainments compared to the native-born population. This can be the case because compulsoryschooling requirements in their countries of origin differ but also because some migrants, especiallyin school-age may have been out of school if they are coming from countries that do not enforcecompulsory education or they had disrupted educational trajectory due to war, refugee situation oras a result of migration process.9 In most HICs, compulsory education is set at completing lower secondary level (i.e. min 8 or 9 yearsof schooling), while in the past it was often set at the level of primary.812

older datasets have been collected and estimates have been made. This was not possible formany OECD countries, see for example the United Kingdom (Case 6) as an extreme case inpoint (see Table 2).Table 2. Overview of available 10 education categories by number of countriesCase1234567Noeducation Incompl.primary Completeprimary N173156183Lowersecondary 180Uppersecondary 185Postsecondary 185No. ofcountries1531485311185Several adjustments were made to the data, similarly to what was implemented for WIC2014 (Bauer et al. 2012). These are documented in the following sub-sections.3.1.23.1.2.1Base-year Data AdjustmentsApplying ISCED 2011Educational categories surveyed in the censuses and surveys tend to be based on nationaleducational programs. Due to the variety of nationally distinct educational systems,UNESCO designed the International Standard Classification of Education (ISCED). In WIC 2014the six educational categories (no formal education, incomplete primary, completedprimary, completed lower secondary, completed upper secondary and post-secondary)were based on ISCED 1997 (UNESCO Institute for Statistics 1997) as ISCED mappings werenot available for the 2011 revision (UNESCO 2012). WIC 2018 is following ISCED 2011,whereby the changes concern the post-secondary category. As mentioned before, wedisaggregated the post-secondary education category for a limited number of countries(60). The allocation procedure is detailed in Annex B. The population of these countriesdivided in eight categories was projected to 2100, but not back projected for lack of historicaldata. Therefore, in the reconstruction, we us the same six education categories as in the WIC2014 dataset (seeTable 3).Available categories include cases where data were readily available from the primary source aswell as cases where they were estimated using additional/alternative data sources. A countryspecific overview about available education categories can be found in Annex D as well as in thesupplementary data documentation files.1013

Table 3. The WIC 2014 educational attainment categories according to ISCED 1997 and 2011classificationISCED 2011ISCED 1997WIC 201401Early imary1Primary2Lower secondary23Upper secondary4e1No e

of educational attainment in some of the major past and present changes whether they are of socio-economic, technological or environmental nature in the mediumto long run. On the supply side, we cannot help noticing that data on educational attainment suffer from several flaws that prevent comparison across years or countries.